Noisy Speech Recognition with Discrete-Mixture HMMs Based on MAP Estimation

In this paper, we develop a novel modeling scheme for discrete-mixture HMMs (DMHMMs) by using maximum a posteriori (MAP) estimation. Also the MAP estimated DMHMMs are used for speech recognition to improve the accuracy under noisy conditions. The DMHMMs were originally proposed to reduce calculation costs in decoding process [1][2]. We propose a new method for MAP estimation of DMHMM parameters to improve trainability. Also we apply the DMHMMs to noisy speech recognition, because models which have discrete probability can represent more complicated shapes and they are expected to be useful for noisy speech recognition. Compared with conventional CHMMs, MAP estimated DMHMMs showed superior recognition performance for noisy speech recognition.